Sei sulla pagina 1di 4

(mt notes)

Whether used during the design


phase of high-performance
craft such as the trimaran
Adastra, or with commercial or
naval vessels, optimization can
result in better understanding
of the design space. Photo
courtesy John Shuttleworth
Yacht Designs Ltd.

Effective Optimization
How numerical optimization tools enable improved design performance

O
BY MATTHEW COLLETTE ptimization is a powerful tool to help naval archi- possible. At the end of this process, the correspond-
tects transform increasingly complex numerical ing design parameters are used as a baseline for further
engineering models into a better understanding design iterations and refinement.
of the design space and improved design perfor- While powerful optimization approaches exist, they
mance. The use of formal numerical optimization within are not a replacement for a coherent overall design strat-
marine design has now become commonplace for a wide egy. Rather, optimization is best viewed as a tool to be
range of vessels, addressing resistance, seakeeping, struc- used within a wider design strategy. Additionally, like any
tural weight, and producibility. This growth in use has engineering tool, optimization is based on a number of
been mirrored by a rapid growth within the optimiza- assumptions and idealization. Successful application of
tion community, with new optimization formulations optimization requires expert judgment in several areas.
and design strategies appearing regularly. Our purpose Judgment is required in selecting appropriate perfor-
here will be to briefly review and contrast several of the mance measures and constraints and turning them into
more established approaches for optimization. objective functions that can be modeled. As the objective
Optimization can be seen as one type of model- functions and constraints provide the only feedback to
based design. In this approach, one or more performance the optimizer itself, it is essential that they include all the
parameters of interest are first identified. These param- relevant design consideration for the problem at hand.
eters become the objective functions of the optimization Constructing the model also requires judgment, as an
process. A numerical model is then built that links design acceptable balance of accuracy and speed is necessary.
parameters describing the design to the designs objective Finally, the selection of the optimizer itself also is
functions. Example models could include computational important. The perfect optimizer does not exist; all opti-
fluid dynamics calculations for resistance, or economic mizers today fundamentally trade efficiency for design
models for vessel performance. Any applicable con- space search power. Additionally, there is no way to
straints, such as cost limits or regulatory compliance guarantee that the globally optimal design solution has
requirements, also must be identified and included in been located for complex design problems. Selecting a
this model. Finally, a suitable optimizer is coupled to correct optimizer for the type of design space and model
this setup; the optimizer strategically modifies the design used is essential to fully maximizing performance.
parameters to deliver the best objective performance We will first explore the formulation of the objective

(18) marine technology July 2014 www.sname.org/sname/mt


functions and model, followed by the types formulation, the preference function on the type of optimizer selected, the con-
of optimizers available today. While were measures the distance in objective func- straint may be incorporated directly into the
exploring these topics sequentially here, tion space from the ideal solution to the objective function formulation in a penalty
decisions made on each topic are in fact achieved solutions on the Pareto front function approach. In this type of approach,
linked in practice. and prefers the solution with the shortest as constraints boundaries are approached or
distance. In a min-max formulation, the violated, the objective functions are artificially
Objective functions preference function seeks to make each worsened so that the infeasible region beyond
and model formulation objective function sacrifice an equal per- the constraint boundary is unpreferred com-
The model linking the problem design centage of their utopian performance. A pared to the feasible domain. While simple
parameters to the objective functions is goal-programming approach may also be and effective, such penalty functions often
a key part of any optimization process. used where individual targets are specified require several tuning runs to develop ade-
Formulation of this model should start by for each objective function and the prefer- quate penalty magnitudes for successful
considering what objective functions are ence function measures how close these optimization. More recent optimizers inde-
relevant to the decisions being made as targets are to obtainment. pendently consider constraints to avoid the
part of the overall design process. In the In addition to the objective function, any iterative nature of penalty functions.
most straightforward case, only a single relevant constraints must also be formulated A final consideration in model formula-
objective, such as cost, profit, or resistance, for the model. These constraints represent tion is to evaluate the need for a surrogate
needs to be considered. However, many limits on types of solutions the optimizer can model, or meta-model of the problem. This
design decisions revolve around finding the consider, such as minimum strength, stabil- technique is widely used when an accu-
correct tradeoff between conflicting design ity, or performance requirements. Depending rate computational model of one or more
objectives. In this case, a multi-objective
formulation becomes necessary. As soon
as there is more than one objective, a deci-
sion has to be made if the model is to support What are USCG/IMO
Horn Requirements?
trade-space exploration or to support deter-
mining a single, balanced, optimal design.
If exploration is the focus of the model,
the individual objective functions will be
solved independently, and a multi-objec-
tive optimizer can be used to find the Pareto
front corresponding to this model. The
Pareto front represents the fundamental
trade space of the problem; for each solu-
tion on the front, it is not possible to further
improve any single objective without sacri-
ficing performance in other objectives.
Alternatively, the various objective
functions can be combined into a single USCG and IMO Regulations require a minimum output for horns on all
vessels over 40 feet long. Surprisingly, many boats are fitted with
preference function. Various approaches
horns and loudhailers that do not meet these requirements.
exist for forming the preference function.
A weighted sum of the (normalized) objec- Kahlenberg Horns are available in air and all electric versions for
tive function values is often the simplest, all types of vessels 12 to over 200 meters L.O.A.. Each shipment
and enables the design team to specify how includes ABS Certificates to USCG and IMO Rules.
important each objective is to the prob-
Have a virtual blast...see and hear them at www.kahlenberg.com.
lem at hand. Geometric definitions related
to the Pareto front can be used to find a USCG/IMO compliant, ABS Type Approved sound signals
design that represents a balance of perfor-
Kahlenberg Industries, Inc. Ph: 920-793-4507
mance between the conflicting objective
Two Rivers, WI USA www.kahlenberg.com
functions. In a nearest-to-the-utopian

www.sname.org/sname/mt July 2014 marine technology (19)


Effective Optimization continued

In a min-max formulation, the preference function


seeks to make each objective function sacrifice an equal
percentage of their utopian performance.

objective function or constraint is too time- than stochastic optimizers, but have lower
consuming or difficult to directly couple to global search power and robustness than
an optimizer. Instead, a rapid surrogate, or the stochastic optimizers.
replacement, computational model is con- Classical optimizers are the most
structed to be used in the optimization well-known optimization techniques. If
process. There are various surrogate model the objective function space is viewed as
forms that are commonly used, including a type of landscape, they attempt to locate
polynomial response surfaces, radial basis the most directly accessible mountaintop
functions, artificial neural networks, and or valley bottom (for example, the max-
Gaussian process models (also known as ima or minima of the objective functions)
Kriging models). Each of these approaches from their current location on the land-
is fundamentally regressiona simplified scape. Classical optimizers can be further
mathematical function is being fitted to divided into two sub-typessearch and
data generated by the time-consuming full gradient methodsdepending on how
model. Then the surrogate model can be to view this landscape. The search meth-
coupled to the optimizer and used in place ods build up a concept of the slope of the
of the full model to drive the optimization landscape space by comparing objective
process. As the surrogate model is only an function evaluations at different locations.
approximation of the full model, it is nor- They do not require direct calculations of
mally necessary to perform checking and the gradients of the objective function at
validation on the surrogate at several points these sampling points. With this data, they
as the optimization process proceeds, to plan which combination of design param-
ensure it is accurate enough for the prob- eters to try next to improve the designs
lem at hand. Normally, several refinements current performance.
of the surrogate model may be necessary in Search methods include the Nelder-Mead
the course of an optimization process. and Powell methods. Gradient methods
require both the objective function evalua-
Optimizer selection tion and the gradient at each point considered;
The model represents only half the opti- however, by intelligently using the gradient
mization problem. An efficient method of information, they can usually converge faster
varying the design parameters to maximize and more precisely than search methods.
the performance of the objective functions Recent examples of gradient-based methods
is also needed. Such a method is referred to include sequential quadratic programming
as an optimizer, and the selection of an opti- methods, which can efficiently solve problems
mizer has a significant impact on the time with complicated objective functions and con-
required and performance of the overall straints. When efficient methods of evaluating
optimization process. There are two broad gradients are available, gradient methods have
categories of optimizers: classical optimiz- been demonstrated to solve complex optimi-
ers and stochastic optimizers. In general, zation problems with 200-400 design variables
classical optimizers are faster to converge in the aerospace community.

(20) marine technology July 2014 www.sname.org/sname/mt


(mt notes)

A potential disadvantage of classical enabled the development of effective multi- of MDO approaches have been proposed
methods is that they try to find the most objective optimizers where each member to date, and many frameworks are under
directly accessible maxima or minima of the of the population is pushed to converge to active research development to further
objective function from their starting point. a different location along the Pareto front. improve their performance.
Thus, they are prone to getting trapped in This results in a comprehensive definition Optimization is a powerful tool for
locally optimal region, and do not have of the Pareto front, which is very useful for use within a design framework. Setting
significant global search power to handle trade studies. up a successful optimization analysis
design spaces where there are many max- A final optimization approach that can requires attention to both the model used
ima or minima. Additionally, non-smooth or be taken for very large problems is that of to inform the optimizer, and the selec-
non-continuous objective functions, as well multi-disciplinary optimization (MDO). tion of the optimizer itself. For each, there
as discrete input variables, can be challeng- In this type of approach, a large optimiza- are a number of well-developed alterna-
ing to these types of formulations. tion problem is broken down into smaller, tives for specific problem types and design
The stochastic optimization methods discipline-based optimization problems. objectives. When such methods are suc-
trade some efficiency and convergence prop- Optimization proceeds by a hierarchy of cessfully matched to the design problem
erties for improved global search power and cooperative optimizerstypically one at hand, optimization can lead to a bet-
robustness in the face of discontinuous or working at the global problem level and ter understanding of the design space
discrete problem formulations. Examples of one working on each discipline. Such a and improved design performance. MT
these methods include genetic algorithms, distributed framework often can often
particle swarm optimization, and simulat- solve large optimization problems faster Matthew Collette is an assistant professor of naval architec-
ing annealing. In general, these methods than a monolithic optimizer. A variety ture and marine engineering at the University of Michigan.
only require the ability to compare to poten-
tial solutions to the optimization problem
and decide which one is better. They make
no assumptions about gradients or search
directions. Instead, they generate new and
hopefully improved solutions by mimicking
processes inspired by natureevolution in
the case of genetic algorithm, swarm or flock
behavior in the case of the particle swarm
optimizer, and the annealing of metals in
the case of simulating annealing. As these
processes involve an element of chance in
nature, these methods typically use quasi-
random numbers generated by a computer
in their search formulations, leading to the
stochastic moniker for this class of algorithm.

Globally optimal
While no method can guarantee completely
effective global search, these algorithms
are fairly tolerant of design spaces where
there are many competing locally opti-
mal points in that the methods often are
able to find most of these points and deter-
mine which one is truly globally optimal.
Additionally, many of these methods are
population-based, in that they search
from a population of individual candidate
solutions. This population formulation has

www.sname.org/sname/mt July 2014 marine technology (21)

Potrebbero piacerti anche